Well, this is going to be my last entry from SuperComputing 2016. There are a number of technical sessions happening tomorrow morning but this evening officially marks the end of the conference itself and I'm flying home early in the afternoon!
As I went through the exhibitors hall to find a quiet spot to write this up a little while ago it was odd, everything that took nearly 2 days to set up on Sunday and Monday was nearly gone, even the carpet that somehow got laid across the entire 55,000 Sq/ft hall in under 24 hours was mostly rolled back up again, amazing really.
Anyway, today there were a number of really good technical sessions, this morning saw Preeti Malakar of Argonne National Lab talking about their software library/layer called FOCAL. Effectively they are interested in optimising the transfer of data in situations where a simulation requires post-processing (or analysis as they called it) to glean useful informaton. Their examples were mainly from Molecular Dynamics performed with LLAMPS but their work applies to anything with the goal being to find the optimal frequency to transfer data for analysis when applications reside in different parts of a distributed system. They showed some great speed-ups, the general approach and download link for FOCAL can be seen in the following slide:
Later I attended the second of the prestigious ACM Gordon Bell finalist papers. This session was looking at climate modelling at extreme HPC scales. One of the talks came from Chao Yang of the Chinese Academy of Sciences and showed their work on re-designing a fully-implicit solver for non-hydrostatic atmospheric dynamics to work across over 10 million cores on the worlds fastest supercomputer, the Sunway TaihuLight. This is a radically different type of computer which has many "small" cores (a little like the IBM BlueGene/Q but on a much bigger scale).
They have produced both an explicit and fully implicit version of their solver and while the explicit solver scales amazingly well on the hardware it is the fact that the implicit code has been made to scale nearly as well and to the level of over 10M cores that is truly outstanding. I still haven't quite digested what they have done technically but it is described in the following slide (for the record… they won the prize!):
It's been a very interesting week. If I had one comment then it would be that in many cases I feel descriptions of scientific applications has been a little thin on the ground, often covered in a single slide. However as the focus here is the computation I think I can probably let that slide! I have been amazed at the scale of the conference and have been interested to see the current directions that the hardware and software manufaturers are heading in. The term cognitive computing appears to be here to stay…
I thought I'd finish where most who attend seem to start, with a (slightly blurry) ubiquitous shot of the SC'16 logo on the Salt Palace Convention Center's main tower: